Manoel
@manoelribeiro
Computational Social Science, Platforms, GenAI, Moderation
Is there a strong case for AI helping, rather than harming, the accuracy of people's beliefs about contentious topics? In this @NatureMedicine piece (focusing on vaccination), I argue the answer is YES. And it boils down to how LLMs differ from other sources of information.
📄NEW PAPER📄 Ever wondered content people actually pay *attention* to online? Our new research reveals that you likely pay attention to far more varied political content than your likes and shares suggest
🚨 Study alert! We've developed a prototype that enables you to create personalized social media feeds using natural language rules and import them into Bluesky. Join our study to try out the tool, give feedback, and earn a $50 gift card! Sign up: docs.google.com/forms/d/e/1FAI…
Thank you for the shoutout, Kiran! It means a lot to me 🥲 This paper is exciting to us because it shows that a decade-old but simple model cascades method scales effectively with LLM advancements to practically address high-stakes health issues like opioid use disorder myths.
This paper presents some nice methods to scalably understand video content. They find a large prevalence of opiod use myths on youtube. Advances in VLMs are making such great analysis possible, which was not possible even 5 years ago. arxiv.org/abs/2506.00308
Very exciting work out of my lab at @ICepfl, led by @SaiboGeng — zip2zip, a method for allowing LLMs to work directly in compressed token space.
🚀 Excited to share our latest work at ICML 2025 — zip2zip: Inference-Time Adaptive Vocabularies for Language Models via Token Compression! Sessions: 📅 Fri 18 Jul - Tokenization Workshop 📅 Sat 19 Jul - Workshop on Efficient Systems for Foundation Models (Oral 5/145)
Mechanistic interpretability often relies on *interventions* to study how DNNs work. Are these interventions enough to guarantee the features we find are not spurious? No!⚠️ In our new paper, we show many mech int methods implicitly rely on the linear representation hypothesis🧵
Can joining local online communities foster civic engagement and local democracy? Our recent @SocialMedia_Soc paper (led by superstar @marianneaubin) tackles this q with a longitudinal study of 2.8K Redditors across 5 countries. journals.sagepub.com/doi/pdf/10.117… Here's what we found:
I am once again being asked if X (or TikTok, or Facebook...) amplifies content from one political orientation more than others. And I am once again telling you that it's complicated to even define what that means, so you should read this excellent writeup doomscrollingbabel.manoel.xyz/p/does-tiktok-…
ICWSM 2025 best paper honorable mention: "Protection from Evil and Good: The Differential Effects of Page Protection on Wikipedia Article Quality" (Thorsten Ruprechter, @manoelribeiro,@cervisiarius,Denis Helic) doi.org/10.1609/icwsm.… Earlier presentation: mediawiki.org/wiki/Wikimedia…
Best paper award and honorable mentions 🏆 Congrats to all authors! #icwsm
July 4th break in our #AI4Science seminar series. Join us next week for a talk by @ChengleiSi on the epic 2-year experiment evaluating (and executing!) AI-generated scientific ideas. lu.ma/9qq72ebt
A good chunk of the AI-generated content we see right now is paid for by VC money. E.g., Leonardo.ai, a popular image/video creator, offers Veo3 **cheaper than Google**! It is like early-day Uber.
Very cool
100+ filmmakers and celebrities voted for their favorite movies of the 21st century in the NYT poll. Whose tastes are most similar to yours? I made a website that lets you find out: Pick your top 10 movies and see your closest matches.
Are AI scientists already better than human researchers? We recruited 43 PhD students to spend 3 months executing research ideas proposed by an LLM agent vs human experts. Main finding: LLM ideas result in worse projects than human ideas.
New Anthropic Research: Project Vend. We had Claude run a small shop in our office lunchroom. Here’s how it went.
New Anthropic Research: Project Vend. We had Claude run a small shop in our office lunchroom. Here’s how it went.